drivers/net/mlx5/mlx5_flow.c

   1 /*-
   2  *   BSD LICENSE
   3  *
   4  *   Copyright 2016 6WIND S.A.
   5  *   Copyright 2016 Mellanox.
   6  *
   7  *   Redistribution and use in source and binary forms, with or without
   8  *   modification, are permitted provided that the following conditions
   9  *   are met:
  10  *
  11  *     * Redistributions of source code must retain the above copyright
  12  *       notice, this list of conditions and the following disclaimer.
  13  *     * Redistributions in binary form must reproduce the above copyright
  14  *       notice, this list of conditions and the following disclaimer in
  15  *       the documentation and/or other materials provided with the
  16  *       distribution.
  17  *     * Neither the name of 6WIND S.A. nor the names of its
  18  *       contributors may be used to endorse or promote products derived
  19  *       from this software without specific prior written permission.
  20  *
  21  *   THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS
  22  *   "AS IS" AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT
  23  *   LIMITED TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR
  24  *   A PARTICULAR PURPOSE ARE DISCLAIMED. IN NO EVENT SHALL THE COPYRIGHT
  25  *   OWNER OR CONTRIBUTORS BE LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL,
  26  *   SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT
  27  *   LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; LOSS OF USE,
  28  *   DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON ANY
  29  *   THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT
  30  *   (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE
  31  *   OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
  32  */
  33
  34 #include <sys/queue.h>
  35 #include <string.h>
  36
  37 /* Verbs header. */
  38 /* ISO C doesn't support unnamed structs/unions, disabling -pedantic. */
  39 #ifdef PEDANTIC
  40 #pragma GCC diagnostic ignored "-Wpedantic"
  41 #endif
  42 #include <infiniband/verbs.h>
  43 #ifdef PEDANTIC
  44 #pragma GCC diagnostic error "-Wpedantic"
  45 #endif
  46
  47 #include <rte_ethdev.h>
  48 #include <rte_flow.h>
  49 #include <rte_flow_driver.h>
  50 #include <rte_malloc.h>
  51
  52 #include "mlx5.h"
  53 #include "mlx5_prm.h"
  54
  55 static int
  56 mlx5_flow_create_eth(const struct rte_flow_item *item,
  57                      const void *default_mask,
  58                      void *data);
  59
  60 static int
  61 mlx5_flow_create_vlan(const struct rte_flow_item *item,
  62                       const void *default_mask,
  63                       void *data);
  64
  65 static int
  66 mlx5_flow_create_ipv4(const struct rte_flow_item *item,
  67                       const void *default_mask,
  68                       void *data);
  69
  70 static int
  71 mlx5_flow_create_ipv6(const struct rte_flow_item *item,
  72                       const void *default_mask,
  73                       void *data);
  74
  75 static int
  76 mlx5_flow_create_udp(const struct rte_flow_item *item,
  77                      const void *default_mask,
  78                      void *data);
  79
  80 static int
  81 mlx5_flow_create_tcp(const struct rte_flow_item *item,
  82                      const void *default_mask,
  83                      void *data);
  84
  85 static int
  86 mlx5_flow_create_vxlan(const struct rte_flow_item *item,
  87                        const void *default_mask,
  88                        void *data);
  89
  90 struct rte_flow {
  91         LIST_ENTRY(rte_flow) next; /**< Pointer to the next flow structure. */
  92         struct ibv_exp_flow_attr *ibv_attr; /**< Pointer to Verbs attributes. */
  93         struct ibv_exp_rwq_ind_table *ind_table; /**< Indirection table. */
  94         struct ibv_qp *qp; /**< Verbs queue pair. */
  95         struct ibv_exp_flow *ibv_flow; /**< Verbs flow. */
  96         struct ibv_exp_wq *wq; /**< Verbs work queue. */
  97         struct ibv_cq *cq; /**< Verbs completion queue. */
  98         struct rxq *(*rxqs)[]; /**< Pointer to the queues array. */
  99         uint16_t rxqs_n; /**< Number of queues in this flow, 0 if drop queue. */
 100         uint32_t mark:1; /**< Set if the flow is marked. */
 101         uint32_t drop:1; /**< Drop queue. */
 102 };
 103
 104 /** Static initializer for items. */
 105 #define ITEMS(...) \
 106         (const enum rte_flow_item_type []){ \
 107                 __VA_ARGS__, RTE_FLOW_ITEM_TYPE_END, \
 108         }
 109
 110 /** Structure to generate a simple graph of layers supported by the NIC. */
 111 struct mlx5_flow_items {
 112         /** List of possible actions for these items. */
 113         const enum rte_flow_action_type *const actions;
 114         /** Bit-masks corresponding to the possibilities for the item. */
 115         const void *mask;
 116         /**
 117          * Default bit-masks to use when item->mask is not provided. When
 118          * \default_mask is also NULL, the full supported bit-mask (\mask) is
 119          * used instead.
 120          */
 121         const void *default_mask;
 122         /** Bit-masks size in bytes. */
 123         const unsigned int mask_sz;
 124         /**
 125          * Conversion function from rte_flow to NIC specific flow.
 126          *
 127          * @param item
 128          *   rte_flow item to convert.
 129          * @param default_mask
 130          *   Default bit-masks to use when item->mask is not provided.
 131          * @param data
 132          *   Internal structure to store the conversion.
 133          *
 134          * @return
 135          *   0 on success, negative value otherwise.
 136          */
 137         int (*convert)(const struct rte_flow_item *item,
 138                        const void *default_mask,
 139                        void *data);
 140         /** Size in bytes of the destination structure. */
 141         const unsigned int dst_sz;
 142         /** List of possible following items.  */
 143         const enum rte_flow_item_type *const items;
 144 };
 145
 146 /** Valid action for this PMD. */
 147 static const enum rte_flow_action_type valid_actions[] = {
 148         RTE_FLOW_ACTION_TYPE_DROP,
 149         RTE_FLOW_ACTION_TYPE_QUEUE,
 150         RTE_FLOW_ACTION_TYPE_MARK,
 151         RTE_FLOW_ACTION_TYPE_FLAG,
 152         RTE_FLOW_ACTION_TYPE_END,
 153 };
 154
 155 /** Graph of supported items and associated actions. */
 156 static const struct mlx5_flow_items mlx5_flow_items[] = {
 157         [RTE_FLOW_ITEM_TYPE_END] = {
 158                 .items = ITEMS(RTE_FLOW_ITEM_TYPE_ETH,
 159                                RTE_FLOW_ITEM_TYPE_VXLAN),
 160         },
 161         [RTE_FLOW_ITEM_TYPE_ETH] = {
 162                 .items = ITEMS(RTE_FLOW_ITEM_TYPE_VLAN,
 163                                RTE_FLOW_ITEM_TYPE_IPV4,
 164                                RTE_FLOW_ITEM_TYPE_IPV6),
 165                 .actions = valid_actions,
 166                 .mask = &(const struct rte_flow_item_eth){
 167                         .dst.addr_bytes = "\xff\xff\xff\xff\xff\xff",
 168                         .src.addr_bytes = "\xff\xff\xff\xff\xff\xff",
 169                         .type = -1,
 170                 },
 171                 .default_mask = &rte_flow_item_eth_mask,
 172                 .mask_sz = sizeof(struct rte_flow_item_eth),
 173                 .convert = mlx5_flow_create_eth,
 174                 .dst_sz = sizeof(struct ibv_exp_flow_spec_eth),
 175         },
 176         [RTE_FLOW_ITEM_TYPE_VLAN] = {
 177                 .items = ITEMS(RTE_FLOW_ITEM_TYPE_IPV4,
 178                                RTE_FLOW_ITEM_TYPE_IPV6),
 179                 .actions = valid_actions,
 180                 .mask = &(const struct rte_flow_item_vlan){
 181                         .tci = -1,
 182                 },
 183                 .default_mask = &rte_flow_item_vlan_mask,
 184                 .mask_sz = sizeof(struct rte_flow_item_vlan),
 185                 .convert = mlx5_flow_create_vlan,
 186                 .dst_sz = 0,
 187         },
 188         [RTE_FLOW_ITEM_TYPE_IPV4] = {
 189                 .items = ITEMS(RTE_FLOW_ITEM_TYPE_UDP,
 190                                RTE_FLOW_ITEM_TYPE_TCP),
 191                 .actions = valid_actions,
 192                 .mask = &(const struct rte_flow_item_ipv4){
 193                         .hdr = {
 194                                 .src_addr = -1,
 195                                 .dst_addr = -1,
 196                                 .type_of_service = -1,
 197                                 .next_proto_id = -1,
 198                         },
 199                 },
 200                 .default_mask = &rte_flow_item_ipv4_mask,
 201                 .mask_sz = sizeof(struct rte_flow_item_ipv4),
 202                 .convert = mlx5_flow_create_ipv4,
 203                 .dst_sz = sizeof(struct ibv_exp_flow_spec_ipv4_ext),
 204         },
 205         [RTE_FLOW_ITEM_TYPE_IPV6] = {
 206                 .items = ITEMS(RTE_FLOW_ITEM_TYPE_UDP,
 207                                RTE_FLOW_ITEM_TYPE_TCP),
 208                 .actions = valid_actions,
 209                 .mask = &(const struct rte_flow_item_ipv6){
 210                         .hdr = {
 211                                 .src_addr = {
 212                                         0xff, 0xff, 0xff, 0xff,
 213                                         0xff, 0xff, 0xff, 0xff,
 214                                         0xff, 0xff, 0xff, 0xff,
 215                                         0xff, 0xff, 0xff, 0xff,
 216                                 },
 217                                 .dst_addr = {
 218                                         0xff, 0xff, 0xff, 0xff,
 219                                         0xff, 0xff, 0xff, 0xff,
 220                                         0xff, 0xff, 0xff, 0xff,
 221                                         0xff, 0xff, 0xff, 0xff,
 222                                 },
 223                                 .vtc_flow = -1,
 224                                 .proto = -1,
 225                                 .hop_limits = -1,
 226                         },
 227                 },
 228                 .default_mask = &rte_flow_item_ipv6_mask,
 229                 .mask_sz = sizeof(struct rte_flow_item_ipv6),
 230                 .convert = mlx5_flow_create_ipv6,
 231                 .dst_sz = sizeof(struct ibv_exp_flow_spec_ipv6_ext),
 232         },
 233         [RTE_FLOW_ITEM_TYPE_UDP] = {
 234                 .items = ITEMS(RTE_FLOW_ITEM_TYPE_VXLAN),
 235                 .actions = valid_actions,
 236                 .mask = &(const struct rte_flow_item_udp){
 237                         .hdr = {
 238                                 .src_port = -1,
 239                                 .dst_port = -1,
 240                         },
 241                 },
 242                 .default_mask = &rte_flow_item_udp_mask,
 243                 .mask_sz = sizeof(struct rte_flow_item_udp),
 244                 .convert = mlx5_flow_create_udp,
 245                 .dst_sz = sizeof(struct ibv_exp_flow_spec_tcp_udp),
 246         },
 247         [RTE_FLOW_ITEM_TYPE_TCP] = {
 248                 .actions = valid_actions,
 249                 .mask = &(const struct rte_flow_item_tcp){
 250                         .hdr = {
 251                                 .src_port = -1,
 252                                 .dst_port = -1,
 253                         },
 254                 },
 255                 .default_mask = &rte_flow_item_tcp_mask,
 256                 .mask_sz = sizeof(struct rte_flow_item_tcp),
 257                 .convert = mlx5_flow_create_tcp,
 258                 .dst_sz = sizeof(struct ibv_exp_flow_spec_tcp_udp),
 259         },
 260         [RTE_FLOW_ITEM_TYPE_VXLAN] = {
 261                 .items = ITEMS(RTE_FLOW_ITEM_TYPE_ETH),
 262                 .actions = valid_actions,
 263                 .mask = &(const struct rte_flow_item_vxlan){
 264                         .vni = "\xff\xff\xff",
 265                 },
 266                 .default_mask = &rte_flow_item_vxlan_mask,
 267                 .mask_sz = sizeof(struct rte_flow_item_vxlan),
 268                 .convert = mlx5_flow_create_vxlan,
 269                 .dst_sz = sizeof(struct ibv_exp_flow_spec_tunnel),
 270         },
 271 };
 272
 273 /** Structure to pass to the conversion function. */
 274 struct mlx5_flow {
 275         struct ibv_exp_flow_attr *ibv_attr; /**< Verbs attribute. */
 276         unsigned int offset; /**< Offset in bytes in the ibv_attr buffer. */
 277         uint32_t inner; /**< Set once VXLAN is encountered. */
 278 };
 279
 280 struct mlx5_flow_action {
 281         uint32_t queue:1; /**< Target is a receive queue. */
 282         uint32_t drop:1; /**< Target is a drop queue. */
 283         uint32_t mark:1; /**< Mark is present in the flow. */
 284         uint32_t mark_id; /**< Mark identifier. */
 285         uint16_t queues[RTE_MAX_QUEUES_PER_PORT]; /**< Queues indexes to use. */
 286         uint16_t queues_n; /**< Number of entries in queue[]. */
 287 };
 288
 289 /**
 290  * Check support for a given item.
 291  *
 292  * @param item[in]
 293  *   Item specification.
 294  * @param mask[in]
 295  *   Bit-masks covering supported fields to compare with spec, last and mask in
 296  *   \item.
 297  * @param size
 298  *   Bit-Mask size in bytes.
 299  *
 300  * @return
 301  *   0 on success.
 302  */
 303 static int
 304 mlx5_flow_item_validate(const struct rte_flow_item *item,
 305                         const uint8_t *mask, unsigned int size)
 306 {
 307         int ret = 0;
 308
 309         if (!item->spec && (item->mask || item->last))
 310                 return -1;
 311         if (item->spec && !item->mask) {
 312                 unsigned int i;
 313                 const uint8_t *spec = item->spec;
 314
 315                 for (i = 0; i < size; ++i)
 316                         if ((spec[i] | mask[i]) != mask[i])
 317                                 return -1;
 318         }
 319         if (item->last && !item->mask) {
 320                 unsigned int i;
 321                 const uint8_t *spec = item->last;
 322
 323                 for (i = 0; i < size; ++i)
 324                         if ((spec[i] | mask[i]) != mask[i])
 325                                 return -1;
 326         }
 327         if (item->mask) {
 328                 unsigned int i;
 329                 const uint8_t *spec = item->mask;
 330
 331                 for (i = 0; i < size; ++i)
 332                         if ((spec[i] | mask[i]) != mask[i])
 333                                 return -1;
 334         }
 335         if (item->spec && item->last) {
 336                 uint8_t spec[size];
 337                 uint8_t last[size];
 338                 const uint8_t *apply = mask;
 339                 unsigned int i;
 340
 341                 if (item->mask)
 342                         apply = item->mask;
 343                 for (i = 0; i < size; ++i) {
 344                         spec[i] = ((const uint8_t *)item->spec)[i] & apply[i];
 345                         last[i] = ((const uint8_t *)item->last)[i] & apply[i];
 346                 }
 347                 ret = memcmp(spec, last, size);
 348         }
 349         return ret;
 350 }
 351
 352 /**
 353  * Validate a flow supported by the NIC.
 354  *
 355  * @param priv
 356  *   Pointer to private structure.
 357  * @param[in] attr
 358  *   Flow rule attributes.
 359  * @param[in] pattern
 360  *   Pattern specification (list terminated by the END pattern item).
 361  * @param[in] actions
 362  *   Associated actions (list terminated by the END action).
 363  * @param[out] error
 364  *   Perform verbose error reporting if not NULL.
 365  * @param[in, out] flow
 366  *   Flow structure to update.
 367  *
 368  * @return
 369  *   0 on success, a negative errno value otherwise and rte_errno is set.
 370  */
 371 static int
 372 priv_flow_validate(struct priv *priv,
 373                    const struct rte_flow_attr *attr,
 374                    const struct rte_flow_item items[],
 375                    const struct rte_flow_action actions[],
 376                    struct rte_flow_error *error,
 377                    struct mlx5_flow *flow)
 378 {
 379         const struct mlx5_flow_items *cur_item = mlx5_flow_items;
 380         struct mlx5_flow_action action = {
 381                 .queue = 0,
 382                 .drop = 0,
 383                 .mark = 0,
 384         };
 385
 386         (void)priv;
 387         if (attr->group) {
 388                 rte_flow_error_set(error, ENOTSUP,
 389                                    RTE_FLOW_ERROR_TYPE_ATTR_GROUP,
 390                                    NULL,
 391                                    "groups are not supported");
 392                 return -rte_errno;
 393         }
 394         if (attr->priority) {
 395                 rte_flow_error_set(error, ENOTSUP,
 396                                    RTE_FLOW_ERROR_TYPE_ATTR_PRIORITY,
 397                                    NULL,
 398                                    "priorities are not supported");
 399                 return -rte_errno;
 400         }
 401         if (attr->egress) {
 402                 rte_flow_error_set(error, ENOTSUP,
 403                                    RTE_FLOW_ERROR_TYPE_ATTR_EGRESS,
 404                                    NULL,
 405                                    "egress is not supported");
 406                 return -rte_errno;
 407         }
 408         if (!attr->ingress) {
 409                 rte_flow_error_set(error, ENOTSUP,
 410                                    RTE_FLOW_ERROR_TYPE_ATTR_INGRESS,
 411                                    NULL,
 412                                    "only ingress is supported");
 413                 return -rte_errno;
 414         }
 415         for (; items->type != RTE_FLOW_ITEM_TYPE_END; ++items) {
 416                 const struct mlx5_flow_items *token = NULL;
 417                 unsigned int i;
 418                 int err;
 419
 420                 if (items->type == RTE_FLOW_ITEM_TYPE_VOID)
 421                         continue;
 422                 for (i = 0;
 423                      cur_item->items &&
 424                      cur_item->items[i] != RTE_FLOW_ITEM_TYPE_END;
 425                      ++i) {
 426                         if (cur_item->items[i] == items->type) {
 427                                 token = &mlx5_flow_items[items->type];
 428                                 break;
 429                         }
 430                 }
 431                 if (!token)
 432                         goto exit_item_not_supported;
 433                 cur_item = token;
 434                 err = mlx5_flow_item_validate(items,
 435                                               (const uint8_t *)cur_item->mask,
 436                                               cur_item->mask_sz);
 437                 if (err)
 438                         goto exit_item_not_supported;
 439                 if (flow->ibv_attr && cur_item->convert) {
 440                         err = cur_item->convert(items,
 441                                                 (cur_item->default_mask ?
 442                                                  cur_item->default_mask :
 443                                                  cur_item->mask),
 444                                                 flow);
 445                         if (err)
 446                                 goto exit_item_not_supported;
 447                 }
 448                 flow->offset += cur_item->dst_sz;
 449         }
 450         for (; actions->type != RTE_FLOW_ACTION_TYPE_END; ++actions) {
 451                 if (actions->type == RTE_FLOW_ACTION_TYPE_VOID) {
 452                         continue;
 453                 } else if (actions->type == RTE_FLOW_ACTION_TYPE_DROP) {
 454                         action.drop = 1;
 455                 } else if (actions->type == RTE_FLOW_ACTION_TYPE_QUEUE) {
 456                         const struct rte_flow_action_queue *queue =
 457                                 (const struct rte_flow_action_queue *)
 458                                 actions->conf;
 459
 460                         if (!queue || (queue->index > (priv->rxqs_n - 1)))
 461                                 goto exit_action_not_supported;
 462                         action.queue = 1;
 463                 } else if (actions->type == RTE_FLOW_ACTION_TYPE_MARK) {
 464                         const struct rte_flow_action_mark *mark =
 465                                 (const struct rte_flow_action_mark *)
 466                                 actions->conf;
 467
 468                         if (!mark) {
 469                                 rte_flow_error_set(error, EINVAL,
 470                                                    RTE_FLOW_ERROR_TYPE_ACTION,
 471                                                    actions,
 472                                                    "mark must be defined");
 473                                 return -rte_errno;
 474                         } else if (mark->id >= MLX5_FLOW_MARK_MAX) {
 475                                 rte_flow_error_set(error, ENOTSUP,
 476                                                    RTE_FLOW_ERROR_TYPE_ACTION,
 477                                                    actions,
 478                                                    "mark must be between 0"
 479                                                    " and 16777199");
 480                                 return -rte_errno;
 481                         }
 482                         action.mark = 1;
 483                 } else if (actions->type == RTE_FLOW_ACTION_TYPE_FLAG) {
 484                         action.mark = 1;
 485                 } else {
 486                         goto exit_action_not_supported;
 487                 }
 488         }
 489         if (action.mark && !flow->ibv_attr && !action.drop)
 490                 flow->offset += sizeof(struct ibv_exp_flow_spec_action_tag);
 491         if (!action.queue && !action.drop) {
 492                 rte_flow_error_set(error, ENOTSUP, RTE_FLOW_ERROR_TYPE_HANDLE,
 493                                    NULL, "no valid action");
 494                 return -rte_errno;
 495         }
 496         return 0;
 497 exit_item_not_supported:
 498         rte_flow_error_set(error, ENOTSUP, RTE_FLOW_ERROR_TYPE_ITEM,
 499                            items, "item not supported");
 500         return -rte_errno;
 501 exit_action_not_supported:
 502         rte_flow_error_set(error, ENOTSUP, RTE_FLOW_ERROR_TYPE_ACTION,
 503                            actions, "action not supported");
 504         return -rte_errno;
 505 }
 506
 507 /**
 508  * Validate a flow supported by the NIC.
 509  *
 510  * @see rte_flow_validate()
 511  * @see rte_flow_ops
 512  */
 513 int
 514 mlx5_flow_validate(struct rte_eth_dev *dev,
 515                    const struct rte_flow_attr *attr,
 516                    const struct rte_flow_item items[],
 517                    const struct rte_flow_action actions[],
 518                    struct rte_flow_error *error)
 519 {
 520         struct priv *priv = dev->data->dev_private;
 521         int ret;
 522         struct mlx5_flow flow = { .offset = sizeof(struct ibv_exp_flow_attr) };
 523
 524         priv_lock(priv);
 525         ret = priv_flow_validate(priv, attr, items, actions, error, &flow);
 526         priv_unlock(priv);
 527         return ret;
 528 }
 529
 530 /**
 531  * Convert Ethernet item to Verbs specification.
 532  *
 533  * @param item[in]
 534  *   Item specification.
 535  * @param default_mask[in]
 536  *   Default bit-masks to use when item->mask is not provided.
 537  * @param data[in, out]
 538  *   User structure.
 539  */
 540 static int
 541 mlx5_flow_create_eth(const struct rte_flow_item *item,
 542                      const void *default_mask,
 543                      void *data)
 544 {
 545         const struct rte_flow_item_eth *spec = item->spec;
 546         const struct rte_flow_item_eth *mask = item->mask;
 547         struct mlx5_flow *flow = (struct mlx5_flow *)data;
 548         struct ibv_exp_flow_spec_eth *eth;
 549         const unsigned int eth_size = sizeof(struct ibv_exp_flow_spec_eth);
 550         unsigned int i;
 551
 552         ++flow->ibv_attr->num_of_specs;
 553         flow->ibv_attr->priority = 2;
 554         eth = (void *)((uintptr_t)flow->ibv_attr + flow->offset);
 555         *eth = (struct ibv_exp_flow_spec_eth) {
 556                 .type = flow->inner | IBV_EXP_FLOW_SPEC_ETH,
 557                 .size = eth_size,
 558         };
 559         if (!spec)
 560                 return 0;
 561         if (!mask)
 562                 mask = default_mask;
 563         memcpy(eth->val.dst_mac, spec->dst.addr_bytes, ETHER_ADDR_LEN);
 564         memcpy(eth->val.src_mac, spec->src.addr_bytes, ETHER_ADDR_LEN);
 565         eth->val.ether_type = spec->type;
 566         memcpy(eth->mask.dst_mac, mask->dst.addr_bytes, ETHER_ADDR_LEN);
 567         memcpy(eth->mask.src_mac, mask->src.addr_bytes, ETHER_ADDR_LEN);
 568         eth->mask.ether_type = mask->type;
 569         /* Remove unwanted bits from values. */
 570         for (i = 0; i < ETHER_ADDR_LEN; ++i) {
 571                 eth->val.dst_mac[i] &= eth->mask.dst_mac[i];
 572                 eth->val.src_mac[i] &= eth->mask.src_mac[i];
 573         }
 574         eth->val.ether_type &= eth->mask.ether_type;
 575         return 0;
 576 }
 577
 578 /**
 579  * Convert VLAN item to Verbs specification.
 580  *
 581  * @param item[in]
 582  *   Item specification.
 583  * @param default_mask[in]
 584  *   Default bit-masks to use when item->mask is not provided.
 585  * @param data[in, out]
 586  *   User structure.
 587  */
 588 static int
 589 mlx5_flow_create_vlan(const struct rte_flow_item *item,
 590                       const void *default_mask,
 591                       void *data)
 592 {
 593         const struct rte_flow_item_vlan *spec = item->spec;
 594         const struct rte_flow_item_vlan *mask = item->mask;
 595         struct mlx5_flow *flow = (struct mlx5_flow *)data;
 596         struct ibv_exp_flow_spec_eth *eth;
 597         const unsigned int eth_size = sizeof(struct ibv_exp_flow_spec_eth);
 598
 599         eth = (void *)((uintptr_t)flow->ibv_attr + flow->offset - eth_size);
 600         if (!spec)
 601                 return 0;
 602         if (!mask)
 603                 mask = default_mask;
 604         eth->val.vlan_tag = spec->tci;
 605         eth->mask.vlan_tag = mask->tci;
 606         eth->val.vlan_tag &= eth->mask.vlan_tag;
 607         return 0;
 608 }
 609
 610 /**
 611  * Convert IPv4 item to Verbs specification.
 612  *
 613  * @param item[in]
 614  *   Item specification.
 615  * @param default_mask[in]
 616  *   Default bit-masks to use when item->mask is not provided.
 617  * @param data[in, out]
 618  *   User structure.
 619  */
 620 static int
 621 mlx5_flow_create_ipv4(const struct rte_flow_item *item,
 622                       const void *default_mask,
 623                       void *data)
 624 {
 625         const struct rte_flow_item_ipv4 *spec = item->spec;
 626         const struct rte_flow_item_ipv4 *mask = item->mask;
 627         struct mlx5_flow *flow = (struct mlx5_flow *)data;
 628         struct ibv_exp_flow_spec_ipv4_ext *ipv4;
 629         unsigned int ipv4_size = sizeof(struct ibv_exp_flow_spec_ipv4_ext);
 630
 631         ++flow->ibv_attr->num_of_specs;
 632         flow->ibv_attr->priority = 1;
 633         ipv4 = (void *)((uintptr_t)flow->ibv_attr + flow->offset);
 634         *ipv4 = (struct ibv_exp_flow_spec_ipv4_ext) {
 635                 .type = flow->inner | IBV_EXP_FLOW_SPEC_IPV4_EXT,
 636                 .size = ipv4_size,
 637         };
 638         if (!spec)
 639                 return 0;
 640         if (!mask)
 641                 mask = default_mask;
 642         ipv4->val = (struct ibv_exp_flow_ipv4_ext_filter){
 643                 .src_ip = spec->hdr.src_addr,
 644                 .dst_ip = spec->hdr.dst_addr,
 645                 .proto = spec->hdr.next_proto_id,
 646                 .tos = spec->hdr.type_of_service,
 647         };
 648         ipv4->mask = (struct ibv_exp_flow_ipv4_ext_filter){
 649                 .src_ip = mask->hdr.src_addr,
 650                 .dst_ip = mask->hdr.dst_addr,
 651                 .proto = mask->hdr.next_proto_id,
 652                 .tos = mask->hdr.type_of_service,
 653         };
 654         /* Remove unwanted bits from values. */
 655         ipv4->val.src_ip &= ipv4->mask.src_ip;
 656         ipv4->val.dst_ip &= ipv4->mask.dst_ip;
 657         ipv4->val.proto &= ipv4->mask.proto;
 658         ipv4->val.tos &= ipv4->mask.tos;
 659         return 0;
 660 }
 661
 662 /**
 663  * Convert IPv6 item to Verbs specification.
 664  *
 665  * @param item[in]
 666  *   Item specification.
 667  * @param default_mask[in]
 668  *   Default bit-masks to use when item->mask is not provided.
 669  * @param data[in, out]
 670  *   User structure.
 671  */
 672 static int
 673 mlx5_flow_create_ipv6(const struct rte_flow_item *item,
 674                       const void *default_mask,
 675                       void *data)
 676 {
 677         const struct rte_flow_item_ipv6 *spec = item->spec;
 678         const struct rte_flow_item_ipv6 *mask = item->mask;
 679         struct mlx5_flow *flow = (struct mlx5_flow *)data;
 680         struct ibv_exp_flow_spec_ipv6_ext *ipv6;
 681         unsigned int ipv6_size = sizeof(struct ibv_exp_flow_spec_ipv6_ext);
 682
 683         ++flow->ibv_attr->num_of_specs;
 684         flow->ibv_attr->priority = 1;
 685         ipv6 = (void *)((uintptr_t)flow->ibv_attr + flow->offset);
 686         *ipv6 = (struct ibv_exp_flow_spec_ipv6_ext) {
 687                 .type = flow->inner | IBV_EXP_FLOW_SPEC_IPV6_EXT,
 688                 .size = ipv6_size,
 689         };
 690         if (!spec)
 691                 return 0;
 692         if (!mask)
 693                 mask = default_mask;
 694         memcpy(ipv6->val.src_ip, spec->hdr.src_addr,
 695                RTE_DIM(ipv6->val.src_ip));
 696         memcpy(ipv6->val.dst_ip, spec->hdr.dst_addr,
 697                RTE_DIM(ipv6->val.dst_ip));
 698         memcpy(ipv6->mask.src_ip, mask->hdr.src_addr,
 699                RTE_DIM(ipv6->mask.src_ip));
 700         memcpy(ipv6->mask.dst_ip, mask->hdr.dst_addr,
 701                RTE_DIM(ipv6->mask.dst_ip));
 702         ipv6->mask.flow_label = mask->hdr.vtc_flow;
 703         ipv6->mask.next_hdr = mask->hdr.proto;
 704         ipv6->mask.hop_limit = mask->hdr.hop_limits;
 705         ipv6->val.flow_label &= ipv6->mask.flow_label;
 706         ipv6->val.next_hdr &= ipv6->mask.next_hdr;
 707         ipv6->val.hop_limit &= ipv6->mask.hop_limit;
 708         return 0;
 709 }
 710
 711 /**
 712  * Convert UDP item to Verbs specification.
 713  *
 714  * @param item[in]
 715  *   Item specification.
 716  * @param default_mask[in]
 717  *   Default bit-masks to use when item->mask is not provided.
 718  * @param data[in, out]
 719  *   User structure.
 720  */
 721 static int
 722 mlx5_flow_create_udp(const struct rte_flow_item *item,
 723                      const void *default_mask,
 724                      void *data)
 725 {
 726         const struct rte_flow_item_udp *spec = item->spec;
 727         const struct rte_flow_item_udp *mask = item->mask;
 728         struct mlx5_flow *flow = (struct mlx5_flow *)data;
 729         struct ibv_exp_flow_spec_tcp_udp *udp;
 730         unsigned int udp_size = sizeof(struct ibv_exp_flow_spec_tcp_udp);
 731
 732         ++flow->ibv_attr->num_of_specs;
 733         flow->ibv_attr->priority = 0;
 734         udp = (void *)((uintptr_t)flow->ibv_attr + flow->offset);
 735         *udp = (struct ibv_exp_flow_spec_tcp_udp) {
 736                 .type = flow->inner | IBV_EXP_FLOW_SPEC_UDP,
 737                 .size = udp_size,
 738         };
 739         if (!spec)
 740                 return 0;
 741         if (!mask)
 742                 mask = default_mask;
 743         udp->val.dst_port = spec->hdr.dst_port;
 744         udp->val.src_port = spec->hdr.src_port;
 745         udp->mask.dst_port = mask->hdr.dst_port;
 746         udp->mask.src_port = mask->hdr.src_port;
 747         /* Remove unwanted bits from values. */
 748         udp->val.src_port &= udp->mask.src_port;
 749         udp->val.dst_port &= udp->mask.dst_port;
 750         return 0;
 751 }
 752
 753 /**
 754  * Convert TCP item to Verbs specification.
 755  *
 756  * @param item[in]
 757  *   Item specification.
 758  * @param default_mask[in]
 759  *   Default bit-masks to use when item->mask is not provided.
 760  * @param data[in, out]
 761  *   User structure.
 762  */
 763 static int
 764 mlx5_flow_create_tcp(const struct rte_flow_item *item,
 765                      const void *default_mask,
 766                      void *data)
 767 {
 768         const struct rte_flow_item_tcp *spec = item->spec;
 769         const struct rte_flow_item_tcp *mask = item->mask;
 770         struct mlx5_flow *flow = (struct mlx5_flow *)data;
 771         struct ibv_exp_flow_spec_tcp_udp *tcp;
 772         unsigned int tcp_size = sizeof(struct ibv_exp_flow_spec_tcp_udp);
 773
 774         ++flow->ibv_attr->num_of_specs;
 775         flow->ibv_attr->priority = 0;
 776         tcp = (void *)((uintptr_t)flow->ibv_attr + flow->offset);
 777         *tcp = (struct ibv_exp_flow_spec_tcp_udp) {
 778                 .type = flow->inner | IBV_EXP_FLOW_SPEC_TCP,
 779                 .size = tcp_size,
 780         };
 781         if (!spec)
 782                 return 0;
 783         if (!mask)
 784                 mask = default_mask;
 785         tcp->val.dst_port = spec->hdr.dst_port;
 786         tcp->val.src_port = spec->hdr.src_port;
 787         tcp->mask.dst_port = mask->hdr.dst_port;
 788         tcp->mask.src_port = mask->hdr.src_port;
 789         /* Remove unwanted bits from values. */
 790         tcp->val.src_port &= tcp->mask.src_port;
 791         tcp->val.dst_port &= tcp->mask.dst_port;
 792         return 0;
 793 }
 794
 795 /**
 796  * Convert VXLAN item to Verbs specification.
 797  *
 798  * @param item[in]
 799  *   Item specification.
 800  * @param default_mask[in]
 801  *   Default bit-masks to use when item->mask is not provided.
 802  * @param data[in, out]
 803  *   User structure.
 804  */
 805 static int
 806 mlx5_flow_create_vxlan(const struct rte_flow_item *item,
 807                        const void *default_mask,
 808                        void *data)
 809 {
 810         const struct rte_flow_item_vxlan *spec = item->spec;
 811         const struct rte_flow_item_vxlan *mask = item->mask;
 812         struct mlx5_flow *flow = (struct mlx5_flow *)data;
 813         struct ibv_exp_flow_spec_tunnel *vxlan;
 814         unsigned int size = sizeof(struct ibv_exp_flow_spec_tunnel);
 815         union vni {
 816                 uint32_t vlan_id;
 817                 uint8_t vni[4];
 818         } id;
 819
 820         ++flow->ibv_attr->num_of_specs;
 821         flow->ibv_attr->priority = 0;
 822         id.vni[0] = 0;
 823         vxlan = (void *)((uintptr_t)flow->ibv_attr + flow->offset);
 824         *vxlan = (struct ibv_exp_flow_spec_tunnel) {
 825                 .type = flow->inner | IBV_EXP_FLOW_SPEC_VXLAN_TUNNEL,
 826                 .size = size,
 827         };
 828         flow->inner = IBV_EXP_FLOW_SPEC_INNER;
 829         if (!spec)
 830                 return 0;
 831         if (!mask)
 832                 mask = default_mask;
 833         memcpy(&id.vni[1], spec->vni, 3);
 834         vxlan->val.tunnel_id = id.vlan_id;
 835         memcpy(&id.vni[1], mask->vni, 3);
 836         vxlan->mask.tunnel_id = id.vlan_id;
 837         /* Remove unwanted bits from values. */
 838         vxlan->val.tunnel_id &= vxlan->mask.tunnel_id;
 839         return 0;
 840 }
 841
 842 /**
 843  * Convert mark/flag action to Verbs specification.
 844  *
 845  * @param flow
 846  *   Pointer to MLX5 flow structure.
 847  * @param mark_id
 848  *   Mark identifier.
 849  */
 850 static int
 851 mlx5_flow_create_flag_mark(struct mlx5_flow *flow, uint32_t mark_id)
 852 {
 853         struct ibv_exp_flow_spec_action_tag *tag;
 854         unsigned int size = sizeof(struct ibv_exp_flow_spec_action_tag);
 855
 856         tag = (void *)((uintptr_t)flow->ibv_attr + flow->offset);
 857         *tag = (struct ibv_exp_flow_spec_action_tag){
 858                 .type = IBV_EXP_FLOW_SPEC_ACTION_TAG,
 859                 .size = size,
 860                 .tag_id = mlx5_flow_mark_set(mark_id),
 861         };
 862         ++flow->ibv_attr->num_of_specs;
 863         return 0;
 864 }
 865
 866 /**
 867  * Complete flow rule creation with a drop queue.
 868  *
 869  * @param priv
 870  *   Pointer to private structure.
 871  * @param flow
 872  *   MLX5 flow attributes (filled by mlx5_flow_validate()).
 873  * @param[out] error
 874  *   Perform verbose error reporting if not NULL.
 875  *
 876  * @return
 877  *   A flow if the rule could be created.
 878  */
 879 static struct rte_flow *
 880 priv_flow_create_action_queue_drop(struct priv *priv,
 881                                    struct mlx5_flow *flow,
 882                                    struct rte_flow_error *error)
 883 {
 884         struct rte_flow *rte_flow;
 885
 886         assert(priv->pd);
 887         assert(priv->ctx);
 888         rte_flow = rte_calloc(__func__, 1, sizeof(*rte_flow), 0);
 889         if (!rte_flow) {
 890                 rte_flow_error_set(error, ENOMEM, RTE_FLOW_ERROR_TYPE_HANDLE,
 891                                    NULL, "cannot allocate flow memory");
 892                 return NULL;
 893         }
 894         rte_flow->cq =
 895                 ibv_exp_create_cq(priv->ctx, 1, NULL, NULL, 0,
 896                                   &(struct ibv_exp_cq_init_attr){
 897                                           .comp_mask = 0,
 898                                   });
 899         if (!rte_flow->cq) {
 900                 rte_flow_error_set(error, ENOMEM,
 901                                    RTE_FLOW_ERROR_TYPE_HANDLE,
 902                                    NULL, "cannot allocate CQ");
 903                 goto error;
 904         }
 905         rte_flow->wq = ibv_exp_create_wq(priv->ctx,
 906                                          &(struct ibv_exp_wq_init_attr){
 907                                          .wq_type = IBV_EXP_WQT_RQ,
 908                                          .max_recv_wr = 1,
 909                                          .max_recv_sge = 1,
 910                                          .pd = priv->pd,
 911                                          .cq = rte_flow->cq,
 912                                          });
 913         if (!rte_flow->wq) {
 914                 rte_flow_error_set(error, ENOMEM,
 915                                    RTE_FLOW_ERROR_TYPE_HANDLE,
 916                                    NULL, "cannot allocate WQ");
 917                 goto error;
 918         }
 919         rte_flow->drop = 1;
 920         rte_flow->ibv_attr = flow->ibv_attr;
 921         rte_flow->ind_table = ibv_exp_create_rwq_ind_table(
 922                 priv->ctx,
 923                 &(struct ibv_exp_rwq_ind_table_init_attr){
 924                         .pd = priv->pd,
 925                         .log_ind_tbl_size = 0,
 926                         .ind_tbl = &rte_flow->wq,
 927                         .comp_mask = 0,
 928                 });
 929         if (!rte_flow->ind_table) {
 930                 rte_flow_error_set(error, ENOMEM, RTE_FLOW_ERROR_TYPE_HANDLE,
 931                                    NULL, "cannot allocate indirection table");
 932                 goto error;
 933         }
 934         rte_flow->qp = ibv_exp_create_qp(
 935                 priv->ctx,
 936                 &(struct ibv_exp_qp_init_attr){
 937                         .qp_type = IBV_QPT_RAW_PACKET,
 938                         .comp_mask =
 939                                 IBV_EXP_QP_INIT_ATTR_PD |
 940                                 IBV_EXP_QP_INIT_ATTR_PORT |
 941                                 IBV_EXP_QP_INIT_ATTR_RX_HASH,
 942                         .pd = priv->pd,
 943                         .rx_hash_conf = &(struct ibv_exp_rx_hash_conf){
 944                                 .rx_hash_function =
 945                                         IBV_EXP_RX_HASH_FUNC_TOEPLITZ,
 946                                 .rx_hash_key_len = rss_hash_default_key_len,
 947                                 .rx_hash_key = rss_hash_default_key,
 948                                 .rx_hash_fields_mask = 0,
 949                                 .rwq_ind_tbl = rte_flow->ind_table,
 950                         },
 951                         .port_num = priv->port,
 952                 });
 953         if (!rte_flow->qp) {
 954                 rte_flow_error_set(error, ENOMEM, RTE_FLOW_ERROR_TYPE_HANDLE,
 955                                    NULL, "cannot allocate QP");
 956                 goto error;
 957         }
 958         if (!priv->started)
 959                 return rte_flow;
 960         rte_flow->ibv_flow = ibv_exp_create_flow(rte_flow->qp,
 961                                                  rte_flow->ibv_attr);
 962         if (!rte_flow->ibv_flow) {
 963                 rte_flow_error_set(error, ENOMEM, RTE_FLOW_ERROR_TYPE_HANDLE,
 964                                    NULL, "flow rule creation failure");
 965                 goto error;
 966         }
 967         return rte_flow;
 968 error:
 969         assert(rte_flow);
 970         if (rte_flow->qp)
 971                 ibv_destroy_qp(rte_flow->qp);
 972         if (rte_flow->ind_table)
 973                 ibv_exp_destroy_rwq_ind_table(rte_flow->ind_table);
 974         if (rte_flow->wq)
 975                 ibv_exp_destroy_wq(rte_flow->wq);
 976         if (rte_flow->cq)
 977                 ibv_destroy_cq(rte_flow->cq);
 978         rte_free(rte_flow);
 979         return NULL;
 980 }
 981
 982 /**
 983  * Complete flow rule creation.
 984  *
 985  * @param priv
 986  *   Pointer to private structure.
 987  * @param flow
 988  *   MLX5 flow attributes (filled by mlx5_flow_validate()).
 989  * @param action
 990  *   Target action structure.
 991  * @param[out] error
 992  *   Perform verbose error reporting if not NULL.
 993  *
 994  * @return
 995  *   A flow if the rule could be created.
 996  */
 997 static struct rte_flow *
 998 priv_flow_create_action_queue(struct priv *priv,
 999                               struct mlx5_flow *flow,
1000                               struct mlx5_flow_action *action,
1001                               struct rte_flow_error *error)
1002 {
1003         struct rte_flow *rte_flow;
1004         unsigned int i;
1005         struct ibv_exp_wq *wq[action->queues_n];
1006
1007         assert(priv->pd);
1008         assert(priv->ctx);
1009         assert(!action->drop);
1010         rte_flow = rte_calloc(__func__, 1,
1011                               sizeof(*rte_flow) + sizeof(struct rxq *) *
1012                               action->queues_n, 0);
1013         if (!rte_flow) {
1014                 rte_flow_error_set(error, ENOMEM, RTE_FLOW_ERROR_TYPE_HANDLE,
1015                                    NULL, "cannot allocate flow memory");
1016                 return NULL;
1017         }
1018         rte_flow->rxqs = (struct rxq *(*)[])((uintptr_t)rte_flow +
1019                                              sizeof(struct rxq *) *
1020                                              action->queues_n);
1021         for (i = 0; i < action->queues_n; ++i) {
1022                 struct rxq_ctrl *rxq;
1023
1024                 rxq = container_of((*priv->rxqs)[action->queues[i]],
1025                                    struct rxq_ctrl, rxq);
1026                 wq[i] = rxq->wq;
1027                 (*rte_flow->rxqs)[i] = &rxq->rxq;
1028                 ++rte_flow->rxqs_n;
1029                 rxq->rxq.mark |= action->mark;
1030         }
1031         rte_flow->mark = action->mark;
1032         rte_flow->ibv_attr = flow->ibv_attr;
1033         rte_flow->ind_table = ibv_exp_create_rwq_ind_table(
1034                 priv->ctx,
1035                 &(struct ibv_exp_rwq_ind_table_init_attr){
1036                         .pd = priv->pd,
1037                         .log_ind_tbl_size = 0,
1038                         .ind_tbl = wq,
1039                         .comp_mask = 0,
1040                 });
1041         if (!rte_flow->ind_table) {
1042                 rte_flow_error_set(error, ENOMEM, RTE_FLOW_ERROR_TYPE_HANDLE,
1043                                    NULL, "cannot allocate indirection table");
1044                 goto error;
1045         }
1046         rte_flow->qp = ibv_exp_create_qp(
1047                 priv->ctx,
1048                 &(struct ibv_exp_qp_init_attr){
1049                         .qp_type = IBV_QPT_RAW_PACKET,
1050                         .comp_mask =
1051                                 IBV_EXP_QP_INIT_ATTR_PD |
1052                                 IBV_EXP_QP_INIT_ATTR_PORT |
1053                                 IBV_EXP_QP_INIT_ATTR_RX_HASH,
1054                         .pd = priv->pd,
1055                         .rx_hash_conf = &(struct ibv_exp_rx_hash_conf){
1056                                 .rx_hash_function =
1057                                         IBV_EXP_RX_HASH_FUNC_TOEPLITZ,
1058                                 .rx_hash_key_len = rss_hash_default_key_len,
1059                                 .rx_hash_key = rss_hash_default_key,
1060                                 .rx_hash_fields_mask = 0,
1061                                 .rwq_ind_tbl = rte_flow->ind_table,
1062                         },
1063                         .port_num = priv->port,
1064                 });
1065         if (!rte_flow->qp) {
1066                 rte_flow_error_set(error, ENOMEM, RTE_FLOW_ERROR_TYPE_HANDLE,
1067                                    NULL, "cannot allocate QP");
1068                 goto error;
1069         }
1070         if (!priv->started)
1071                 return rte_flow;
1072         rte_flow->ibv_flow = ibv_exp_create_flow(rte_flow->qp,
1073                                                  rte_flow->ibv_attr);
1074         if (!rte_flow->ibv_flow) {
1075                 rte_flow_error_set(error, ENOMEM, RTE_FLOW_ERROR_TYPE_HANDLE,
1076                                    NULL, "flow rule creation failure");
1077                 goto error;
1078         }
1079         return rte_flow;
1080 error:
1081         assert(rte_flow);
1082         if (rte_flow->qp)
1083                 ibv_destroy_qp(rte_flow->qp);
1084         if (rte_flow->ind_table)
1085                 ibv_exp_destroy_rwq_ind_table(rte_flow->ind_table);
1086         rte_free(rte_flow);
1087         return NULL;
1088 }
1089
1090 /**
1091  * Convert a flow.
1092  *
1093  * @param priv
1094  *   Pointer to private structure.
1095  * @param[in] attr
1096  *   Flow rule attributes.
1097  * @param[in] pattern
1098  *   Pattern specification (list terminated by the END pattern item).
1099  * @param[in] actions
1100  *   Associated actions (list terminated by the END action).
1101  * @param[out] error
1102  *   Perform verbose error reporting if not NULL.
1103  *
1104  * @return
1105  *   A flow on success, NULL otherwise.
1106  */
1107 static struct rte_flow *
1108 priv_flow_create(struct priv *priv,
1109                  const struct rte_flow_attr *attr,
1110                  const struct rte_flow_item items[],
1111                  const struct rte_flow_action actions[],
1112                  struct rte_flow_error *error)
1113 {
1114         struct rte_flow *rte_flow;
1115         struct mlx5_flow_action action;
1116         struct mlx5_flow flow = { .offset = sizeof(struct ibv_exp_flow_attr), };
1117         int err;
1118
1119         err = priv_flow_validate(priv, attr, items, actions, error, &flow);
1120         if (err)
1121                 goto exit;
1122         flow.ibv_attr = rte_malloc(__func__, flow.offset, 0);
1123         flow.offset = sizeof(struct ibv_exp_flow_attr);
1124         if (!flow.ibv_attr) {
1125                 rte_flow_error_set(error, ENOMEM, RTE_FLOW_ERROR_TYPE_HANDLE,
1126                                    NULL, "cannot allocate ibv_attr memory");
1127                 goto exit;
1128         }
1129         *flow.ibv_attr = (struct ibv_exp_flow_attr){
1130                 .type = IBV_EXP_FLOW_ATTR_NORMAL,
1131                 .size = sizeof(struct ibv_exp_flow_attr),
1132                 .priority = attr->priority,
1133                 .num_of_specs = 0,
1134                 .port = 0,
1135                 .flags = 0,
1136                 .reserved = 0,
1137         };
1138         flow.inner = 0;
1139         claim_zero(priv_flow_validate(priv, attr, items, actions,
1140                                       error, &flow));
1141         action = (struct mlx5_flow_action){
1142                 .queue = 0,
1143                 .drop = 0,
1144                 .mark = 0,
1145                 .mark_id = MLX5_FLOW_MARK_DEFAULT,
1146         };
1147         for (; actions->type != RTE_FLOW_ACTION_TYPE_END; ++actions) {
1148                 if (actions->type == RTE_FLOW_ACTION_TYPE_VOID) {
1149                         continue;
1150                 } else if (actions->type == RTE_FLOW_ACTION_TYPE_QUEUE) {
1151                         action.queue = 1;
1152                         action.queues[action.queues_n++] =
1153                                 ((const struct rte_flow_action_queue *)
1154                                  actions->conf)->index;
1155                 } else if (actions->type == RTE_FLOW_ACTION_TYPE_DROP) {
1156                         action.drop = 1;
1157                         action.mark = 0;
1158                 } else if (actions->type == RTE_FLOW_ACTION_TYPE_MARK) {
1159                         const struct rte_flow_action_mark *mark =
1160                                 (const struct rte_flow_action_mark *)
1161                                 actions->conf;
1162
1163                         if (mark)
1164                                 action.mark_id = mark->id;
1165                         action.mark = !action.drop;
1166                 } else if (actions->type == RTE_FLOW_ACTION_TYPE_FLAG) {
1167                         action.mark = 1;
1168                 } else {
1169                         rte_flow_error_set(error, ENOTSUP,
1170                                            RTE_FLOW_ERROR_TYPE_ACTION,
1171                                            actions, "unsupported action");
1172                         goto exit;
1173                 }
1174         }
1175         if (action.mark) {
1176                 mlx5_flow_create_flag_mark(&flow, action.mark_id);
1177                 flow.offset += sizeof(struct ibv_exp_flow_spec_action_tag);
1178         }
1179         if (action.drop)
1180                 rte_flow =
1181                         priv_flow_create_action_queue_drop(priv, &flow, error);
1182         else
1183                 rte_flow = priv_flow_create_action_queue(priv, &flow, &action,
1184                                                          error);
1185         if (!rte_flow)
1186                 goto exit;
1187         return rte_flow;
1188 exit:
1189         rte_free(flow.ibv_attr);
1190         return NULL;
1191 }
1192
1193 /**
1194  * Create a flow.
1195  *
1196  * @see rte_flow_create()
1197  * @see rte_flow_ops
1198  */
1199 struct rte_flow *
1200 mlx5_flow_create(struct rte_eth_dev *dev,
1201                  const struct rte_flow_attr *attr,
1202                  const struct rte_flow_item items[],
1203                  const struct rte_flow_action actions[],
1204                  struct rte_flow_error *error)
1205 {
1206         struct priv *priv = dev->data->dev_private;
1207         struct rte_flow *flow;
1208
1209         priv_lock(priv);
1210         flow = priv_flow_create(priv, attr, items, actions, error);
1211         if (flow) {
1212                 LIST_INSERT_HEAD(&priv->flows, flow, next);
1213                 DEBUG("Flow created %p", (void *)flow);
1214         }
1215         priv_unlock(priv);
1216         return flow;
1217 }
1218
1219 /**
1220  * Destroy a flow.
1221  *
1222  * @param priv
1223  *   Pointer to private structure.
1224  * @param[in] flow
1225  *   Flow to destroy.
1226  */
1227 static void
1228 priv_flow_destroy(struct priv *priv,
1229                   struct rte_flow *flow)
1230 {
1231         (void)priv;
1232         LIST_REMOVE(flow, next);
1233         if (flow->ibv_flow)
1234                 claim_zero(ibv_exp_destroy_flow(flow->ibv_flow));
1235         if (flow->qp)
1236                 claim_zero(ibv_destroy_qp(flow->qp));
1237         if (flow->ind_table)
1238                 claim_zero(ibv_exp_destroy_rwq_ind_table(flow->ind_table));
1239         if (flow->drop && flow->wq)
1240                 claim_zero(ibv_exp_destroy_wq(flow->wq));
1241         if (flow->drop && flow->cq)
1242                 claim_zero(ibv_destroy_cq(flow->cq));
1243         if (flow->mark) {
1244                 struct rte_flow *tmp;
1245                 struct rxq *rxq;
1246                 uint32_t mark_n = 0;
1247                 uint32_t queue_n;
1248
1249                 /*
1250                  * To remove the mark from the queue, the queue must not be
1251                  * present in any other marked flow (RSS or not).
1252                  */
1253                 for (queue_n = 0; queue_n < flow->rxqs_n; ++queue_n) {
1254                         rxq = (*flow->rxqs)[queue_n];
1255                         for (tmp = LIST_FIRST(&priv->flows);
1256                              tmp;
1257                              tmp = LIST_NEXT(tmp, next)) {
1258                                 uint32_t tqueue_n;
1259
1260                                 if (tmp->drop)
1261                                         continue;
1262                                 for (tqueue_n = 0;
1263                                      tqueue_n < tmp->rxqs_n;
1264                                      ++tqueue_n) {
1265                                         struct rxq *trxq;
1266
1267                                         trxq = (*tmp->rxqs)[tqueue_n];
1268                                         if (rxq == trxq)
1269                                                 ++mark_n;
1270                                 }
1271                         }
1272                         rxq->mark = !!mark_n;
1273                 }
1274         }
1275         rte_free(flow->ibv_attr);
1276         DEBUG("Flow destroyed %p", (void *)flow);
1277         rte_free(flow);
1278 }
1279
1280 /**
1281  * Destroy a flow.
1282  *
1283  * @see rte_flow_destroy()
1284  * @see rte_flow_ops
1285  */
1286 int
1287 mlx5_flow_destroy(struct rte_eth_dev *dev,
1288                   struct rte_flow *flow,
1289                   struct rte_flow_error *error)
1290 {
1291         struct priv *priv = dev->data->dev_private;
1292
1293         (void)error;
1294         priv_lock(priv);
1295         priv_flow_destroy(priv, flow);
1296         priv_unlock(priv);
1297         return 0;
1298 }
1299
1300 /**
1301  * Destroy all flows.
1302  *
1303  * @param priv
1304  *   Pointer to private structure.
1305  */
1306 static void
1307 priv_flow_flush(struct priv *priv)
1308 {
1309         while (!LIST_EMPTY(&priv->flows)) {
1310                 struct rte_flow *flow;
1311
1312                 flow = LIST_FIRST(&priv->flows);
1313                 priv_flow_destroy(priv, flow);
1314         }
1315 }
1316
1317 /**
1318  * Destroy all flows.
1319  *
1320  * @see rte_flow_flush()
1321  * @see rte_flow_ops
1322  */
1323 int
1324 mlx5_flow_flush(struct rte_eth_dev *dev,
1325                 struct rte_flow_error *error)
1326 {
1327         struct priv *priv = dev->data->dev_private;
1328
1329         (void)error;
1330         priv_lock(priv);
1331         priv_flow_flush(priv);
1332         priv_unlock(priv);
1333         return 0;
1334 }
1335
1336 /**
1337  * Remove all flows.
1338  *
1339  * Called by dev_stop() to remove all flows.
1340  *
1341  * @param priv
1342  *   Pointer to private structure.
1343  */
1344 void
1345 priv_flow_stop(struct priv *priv)
1346 {
1347         struct rte_flow *flow;
1348
1349         for (flow = LIST_FIRST(&priv->flows);
1350              flow;
1351              flow = LIST_NEXT(flow, next)) {
1352                 claim_zero(ibv_exp_destroy_flow(flow->ibv_flow));
1353                 flow->ibv_flow = NULL;
1354                 if (flow->mark) {
1355                         unsigned int n;
1356
1357                         for (n = 0; n < flow->rxqs_n; ++n)
1358                                 (*flow->rxqs)[n]->mark = 0;
1359                 }
1360                 DEBUG("Flow %p removed", (void *)flow);
1361         }
1362 }
1363
1364 /**
1365  * Add all flows.
1366  *
1367  * @param priv
1368  *   Pointer to private structure.
1369  *
1370  * @return
1371  *   0 on success, a errno value otherwise and rte_errno is set.
1372  */
1373 int
1374 priv_flow_start(struct priv *priv)
1375 {
1376         struct rte_flow *flow;
1377
1378         for (flow = LIST_FIRST(&priv->flows);
1379              flow;
1380              flow = LIST_NEXT(flow, next)) {
1381                 flow->ibv_flow = ibv_exp_create_flow(flow->qp,
1382                                                      flow->ibv_attr);
1383                 if (!flow->ibv_flow) {
1384                         DEBUG("Flow %p cannot be applied", (void *)flow);
1385                         rte_errno = EINVAL;
1386                         return rte_errno;
1387                 }
1388                 DEBUG("Flow %p applied", (void *)flow);
1389                 if (flow->mark) {
1390                         unsigned int n;
1391
1392                         for (n = 0; n < flow->rxqs_n; ++n)
1393                                 (*flow->rxqs)[n]->mark = 1;
1394                 }
1395         }
1396         return 0;
1397 }